What happens when an AI agent cannot complete the task you gave it?
In most products, one of three bad things: it loops, it apologises, or it produces something that looks like a result and is not. ESR.co treats a blocked task as a distinct outcome with its own handling. It classifies why it is blocked into one of five reasons, hands you whatever real work it did manage to produce along with copy-and-paste steps to finish the job by hand, and does not charge Power for an impossible task.
Five reasons a task is blocked, and only three of them are gaps
When ESR cannot finish something, the reason gets classified before anything else happens. The split matters because it decides whether the failure counts as evidence about what to build next.
- Missing integration. You asked for an action against a service ESR has no connection to. Posting to Instagram, running Google Ads.
- Missing credential. The connection exists but is not authorised for your account.
- Not supported. ESR has no capability for this at all.
- Ambiguous. The request was not actionable and needs clarification.
- Transient. An API was down or rate limited, so it should be retried.
Only the first three are logged as demand for new integrations. Ambiguous and transient are deliberately excluded, and the reason is worth stealing if you are building something similar. If unclear requests and temporary outages get counted as unmet demand, the roadmap fills with false signal and you end up building the wrong integration because a service was briefly down.
The manual path
A blocked task still has to leave you better off. If the agent produced genuine work before it hit the wall, that work is handed over as it is, along with numbered steps for finishing the task yourself. Ask for a Google Ads campaign and you get the campaign plan the agent wrote plus the actual sequence of clicks in the Google Ads interface.
There is one rule attached to that, written into the code: the fallback has to contain real work, or say plainly that none could be produced. It is never allowed to be an apology dressed up as a deliverable. That failure mode is common enough to have earned its own guard.
Blocked work is not charged
An impossible task does not consume your Power allowance. That rule is fixed. The genuinely arguable case is the one where the agent did produce usable work on the manual path, since that is real value delivered, and the default there is still not to charge.
Why this is the question to ask about agent reliability
Success rates on agent benchmarks tell you how often a system finishes. They tell you nothing about the shape of the other outcomes, and in day to day use the other outcomes are what determine whether the tool is worth having. A system that finishes 70% of tasks and hands you a usable partial result on the rest is better to work with than one that finishes 80% and silently invents the remainder.
Last reviewed July 26, 2026 by ESR AI CORP. Everything on this page describes what ESR.co ships today. Where a number is given it is read from the product’s own pricing module, not from marketing copy.